EditorNodesPricingBlog

TRELLIS.2 LoRA: Open Source 3D PBR Asset Fine Tuning

November 20, 2026
TRELLIS.2 LoRA: Open Source 3D PBR Asset Fine Tuning

Microsoft Research, TRELLIS.2-4B (MIT License)

Share this post:

TRELLIS.2 LoRA: Open Source 3D PBR Asset Fine Tuning

TRELLIS.2 (microsoft/TRELLIS.2-4B, MIT License) is Microsoft Research's 4 billion parameter flow matching 3D asset generator that turns a single 2D concept still into a textured, relightable 3D mesh with full Physically Based Rendering (PBR) maps, base color, roughness, metallic, and opacity, up to 1536^3 resolution. On November 12, 2026, Hugging Face released an official open source LoRA fine tuning pipeline for TRELLIS.2, enabling VFX studios and indie art departments to adapt the 4B sparse voxel transformer to custom film props, sci fi armor, vehicles, or stylized character geometry without retraining from scratch.

Studio availability. Try the public Hugging Face demo below; TRELLIS.2 is not part of a paid AI FILMS Studio plan (generate concept plates in the image workspace).

TRELLIS.2 4B O-Voxel 3D PBR asset generation gallery

Why O-Voxel and Stage Targeted LoRA Matter for Film VFX

Earlier open 3D generators relied on signed distance fields (SDF) or marching cubes iso surfaces. Those representations sealed hollow cavities shut and melted open surfaces such as torn cloaks, foliage leaves, eyeglasses, and bicycle spokes into thick blobby meshes.

TRELLIS.2 and the November 2026 LoRA release remove both geometry and customization limits:

  1. Field free O-Voxel sparse representation. Instead of forcing every prop into a watertight signed distance shell, TRELLIS.2 encodes assets as field free sparse O-Voxels compressed by a Sparse 3D VAE (SC-VAE) with 16x spatial downsampling. Open cloth, interior cockpits, thin wires, and transparent visors reconstruct cleanly up to 1536^3 effective resolution.
  2. Full PBR material synthesis. Every generated .glb or .usd asset carries separate albedo, roughness, metallic, and opacity channels so lighting artists can drop props straight into Blender, Unreal Engine, or Houdini and relight them under any HDRI dome.
  3. Stage targeted LoRA fine tuning (November 2026). Hugging Face's PEFT recipe targets the sparse structure and shape flow DiT stages while freezing the visual conditioning and texture decoders. Small studio datasets (50 to 300 reference 3D meshes) train specialized geometric LoRAs in hours on a single high VRAM GPU or cloud job instead of multi node cluster runs.
TRELLIS.2 two stage O-Voxel flow matching architecture diagram

TRELLIS.2 two stage flow matching architecture pairing Sparse 3D VAE compression with O-Voxel PBR decoding. Credit: Microsoft Research (MIT License).

Separating sparse structure generation from fine shape and PBR flow matching makes LoRA adaptation surgical.

Art teams can lock their film's exact architectural silhouette or costume proportions by training LoRA adapters strictly on the geometry flow DiT blocks while preserving TRELLIS.2's general PBR lighting priors.

Try TRELLIS.2 Live in the Browser

You can test TRELLIS.2 image to 3D PBR generation directly in Microsoft's official Hugging Face Space below (microsoft-trellis-2.hf.space).

Hardware, Resolution, and License Matrix

Workflow Mode Parameters VRAM Needed Output Resolution License
TRELLIS.2-4B Inference (512^3 – 1024^3) 4B DiT + SC-VAE 16 GB – 24 GB VRAM .glb / .usd with 2K–4K PBR maps MIT License (Commercial OK)
TRELLIS.2-4B High Detail (1536^3) 4B DiT + SC-VAE 24 GB – 40 GB VRAM Hero prop mesh with open surfaces MIT License (Commercial OK)
TRELLIS.2 LoRA Fine Tuning (PEFT) 4B frozen + LoRA ranks (r=16–64) 24 GB – 48 GB VRAM Custom domain LoRA adapter (.safetensors) MIT License (Commercial OK)

Side by side fidelity checks show TRELLIS.2 resolving sharp bevels, translucent glass, metallic specular response, and complex thin structures where older SDF models fuse details together.

Because the weights and fine tuning scripts carry the **MIT license**, studios can ship custom LoRA props into commercial films, games, and LED volume virtual production walls without royalties.

TRELLIS.2 geometric and PBR material fidelity comparison against prior 3D generators

Geometric sharpness and PBR material fidelity comparison in TRELLIS.2-4B. Credit: Microsoft Research (MIT License).

Parameter Reference Table

Parameter Recommended Value Effect on 3D Mesh & LoRA Training
resolution 1024 or 1536 Controls O-Voxel grid density; use 1024 for background set dressing and 1536 for hero close ups.
lora_rank (r) 16 to 32 Balances geometric style capacity against overfitting on small 50–150 asset film prop libraries.
target_stage sparse_structure + shape_flow Focuses LoRA updates on silhouette and topology while keeping general PBR texture priors intact.
texture_size 2048 or 4096 Sets baked UV map size for base color, roughness, metallic, and opacity maps.
cfg_strength 5.0 – 7.5 Tightens adherence to the input concept illustration without clipping metallic reflections.

Local Setup with Python and LoRA Inference

Install microsoft/TRELLIS.2, load the 4B checkpoint, and export a PBR textured .glb ready for Blender or Unreal Engine:

git clone https://github.com/microsoft/TRELLIS.2.git
cd TRELLIS.2
pip install -r requirements.txt
pip install peft diffusers huggingface_hub
import torch
from PIL import Image
from trellis2.pipelines import Trellis2ImageTo3DPipeline

pipe = Trellis2ImageTo3DPipeline.from_pretrained("microsoft/TRELLIS.2-4B")
pipe.cuda()

# Optional: attach a custom film prop LoRA trained with the Nov 2026 PEFT pipeline
# pipe.load_lora_weights("./loras/retro_scifi_props_r32.safetensors")

concept_still = Image.open("hero_prop_helmet_front.png").convert("RGBA")

outputs = pipe.run(
    concept_still,
    seed=108,
    sparse_structure_sampler_params={"steps": 12, "cfg_strength": 6.5},
    slat_sampler_params={"steps": 12, "cfg_strength": 5.0},
)

mesh = outputs["mesh"][0]
mesh.export_glb("hero_prop_helmet_pbr.glb", texture_size=2048)

Filmmaker Production Playbook

  1. Generate clean orthographic concept stills first. Render neutral lit, uncluttered prop stills in FLUX 2 or MiniMax-H3-Image before passing them to TRELLIS.2 so baked shadows do not pollute the albedo map.
  2. Train one show specific LoRA per art department pack. Export 80 to 150 existing kitbash meshes in your film's visual style, run the November 2026 Hugging Face LoRA script on the shape flow stage, and generate dozens of matching background props in minutes.
  3. Preview directly in OpenUSD and Unreal Engine. Export .glb or convert to .usdz so virtual cinematography teams can block real camera lenses around your AI generated props before final video generation.

Related Reading


Sources